Error initializing MCP client for connection on Vercel production build

[kyleledbetter]

Description

ai 4.3.0
next 15.2.0

In local dev this works fine but when I deploy an app with MCP client creation w/ the ai sdk to Vercel I get:

Error initializing MCP client for connection 3cb562fa-f5a3-4111-b76b-83f4275eba8e: Error [MCPClientError]: Connection closed

based on a stdio call trying to run the npx command and getting

npm error enoent Invalid response body while trying to fetch https://registry.npmjs.org/@supabase%2fmcp-server-supabase: ENOENT: no such file or directory, mkdir '/home/sbx_user1051'
npm error enoent This is related to npm not being able to find a file.

I'm using the Supabase MCP with this standard io setup:

{
  "mcpServers": {
    "supabase": {
      "command": "npx",
      "args": [
        "-y",
        "@supabase/mcp-server-supabase@latest",
        "--access-token",
        "<personal-access-token>"
      ]
    }
  }
}

I've got a fairly vanilla Vercel setup for Next.js, not using Fluid Compute

I suspect the serverless nature of a Vercel is the issue, but wondering if other folks have seen this?

I've gotten the same errors using Experimental_StdioMCPTransport and experimental_createMCPClient in an actions.ts as was all my chat route.ts with export const dynamic = 'force-dynamic';

Code example

No response

AI provider

"ai": "^4.3.0", "@ai-sdk/react": "^1.2.6",

Additional context

No response

[nicoalbanese]

Responded on X. stdio is only for local usage, so you'll need to deploy the MCP server and then connect to it via SSE.
Vercel template for MCP servers here: https://vercel.com/templates/other/model-context-protocol-mcp-with-vercel-functions

[kyleledbetter]

Ah that makes perfect sense @nicoalbanese don't know why that didn't hit me. The bug label should definitely be removed, if anything it would be more of a wild feature request to somehow dynamically build client MCP servers on the fly in a serverless env :)

Even though it's probably obvious to many, a documentation enhancement would be to call out explicitly that MCP sdtio clients only run locally for Experimental_StdioMCPTransport in
https://sdk.vercel.ai/cookbook/node/mcp-tools
https://sdk.vercel.ai/docs/reference/ai-sdk-core/mcp-stdio-transport

Related: now i'm seeing Tobias' comment These can be either stdio if running locally in the ai-chatbot PR for mcp also vercel/ai-chatbot#896

[kyleledbetter]

I'll reach out the Supabase team to see when an SSE hosted version of the supabase-mcp is planned